|Publication number||US6981110 B1|
|Application number||US 10/065,340|
|Publication date||Dec 27, 2005|
|Filing date||Oct 6, 2002|
|Priority date||Oct 23, 2001|
|Also published as||US7107402, US7441088, US7444481, US7475200, US7475201, US7478209, US7487304, US7496721, US7506104, US7739452, US20090193216|
|Publication number||065340, 10065340, US 6981110 B1, US 6981110B1, US-B1-6981110, US6981110 B1, US6981110B1|
|Inventors||Stephen Waller Melvin|
|Original Assignee||Stephen Waller Melvin|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (6), Non-Patent Citations (12), Referenced by (39), Classifications (14), Legal Events (6)|
|External Links: USPTO, USPTO Assignment, Espacenet|
The present patent application is a non-provisional application claiming priority to four provisional patent applications as follows: 60/347,781 filed on Oct. 23, 2001; 60/339,077 filed on Oct. 30, 2001; 60/348,217 filed on Nov. 9, 2001; and 60/332,425 filed on Nov. 16, 2001. Each of these priority documents is incorporated in its entirety herein, at least by reference.
Internet networking hardware involves processing of packets of information for many purposes and at many stages in a network. Routers, firewalls, gateways, load balancers and servers all process packets of information in some way. Where in the network the processing takes place (i.e. in the core or close to the edge) has a great deal to do with what types of processing needs to take place and how fast that processing must occur. In general, processing closer to the core takes place faster and involves less work. For example, many core routers perform only layer 2 packet forwarding (i.e. link layer header modification), which can be done with minimal processing overhead. Edge routers, however, typically perform more functions, such as traffic shaping, monitoring, billing and quality of service enforcement. In both situations, the need for processing is constantly evolving, and there is an increasing need to do more at faster rates.
Two key trends are the increase in network speed and the increase in the amount of processing that needs to take place at each stage in the network. Together these trends are forcing packet processing solutions into greater degrees of parallelism.
In the first scenario of
In many cases the workload time is dominated by memory latency due to poor locality of data references and large working set sizes. This means that the limitation on packet throughput is driven by memory throughput, which has tended to increase at a rate even slower than single-threaded processor performance, further driving packet processing solutions into parallel packet processing scenarios.
In the case that all packets can be operated on independently, as shown in the second scenario of
In the more general case that dependencies can arise between packets, a more complicated situation exists. This is shown in the third and fourth scenarios of
Other examples in which packet dependencies can arise would be the updating of traffic management counters and the updating of routing or address translation tables. In the latter case, two packets may be dependent even if they are from completely independent connections if they hash to the same table entry. One packet may want to modify a table entry while another packet is querying the same entry. The fourth scenario in
In these cases where simultaneous processing of packets is required, and where dependencies can exist between packets, it can be complicated to enforce those dependencies. Currently, there are two common approaches to this problem. The first solution is a software solution, where software locks are included in the code to cause dependent packet processing to be delayed until an earlier packet has been completed. These software semaphores are used to lock out subsequent dependent packets from accessing state until the first packet has updated it. The second solution involves hardware, where packet classification hardware serializes all packets that can possibly be dependent. In a multiprocessor, this can involve generating a hash function that sends all packets of the same flow to the same processor, and distributes the load across multiple processors.
Unfortunately, packet processing code is often large and complex and modifying it to incorporate new locking mechanisms is not trivial. Even when such code is relatively simple, guaranteeing that software locks have been correctly programmed for all possible network traffic scenarios can be hard to verify. Furthermore, requiring hardware to enforce sequentiality when it is not needed lowers performance. This is because often the entire packet processing is not dependent such that a partial overlap is possible. The importance of a partial overlap of packet workload can be appreciated by referring to
It is also the case that hardware solutions that group flows for multiprocessors suffer from the problem of guaranteeing that the grouping is relatively uniform over time in order to balance work across the multiple processing elements. The classification of packets to direct them to processing elements is constrained by having to preserve correctness and can't take advantage of a more dynamic load balancing approach.
What is needed is a hardware mechanism to preserve packet dependencies without requiring changes to software and allowing optimal enforcement of dependencies, such that packets are not serialized unless necessary by the overlying application.
The present invention is a hardware mechanism for enforcing packet dependencies without any changes in software. This mechanism has the important advantage that packet dependencies are enforced by hardware only when required and potentially with no discarded work. The software can be written with the view that each packet is handled in order of arrival with no overlap with subsequent packets. This mechanism optimizes performance for the common case that no dependencies are present and doesn't require the hardware to be pre-configured with knowledge of what packet sequences will and will not be dependent.
The present invention is incorporated into a packet processor in which multiple packets are processed simultaneously. This can be implemented using a multiprocessor, a multithreaded processor or a combination of both. The mechanism is not dependent on the type of processor used. It would be useful in any form of packet processor in which some form of simultaneity exists such that more than one packet at a time is being processed.
There are a few requirements placed on the packet processor. The first is that a time stamp or sequence number is associated with each packet as it arrives. The sequence number is used to enforce the sequential processing model. When a packet enters the processor, a new sequence number is generated and that sequence number is associated with the instructions that are processing that packet. The packet processor is required to include this sequence number with all memory operations (reads and writes) performed. An additional requirement is that when a packet enters the processor and when a packet leaves the processor (or when processing ends for that packet), a notification of that event is made to the hardware. The information about when a packet enters and leaves is used to know when write data can be committed to memory.
Finally, the packet processor must have the ability to restart processing for each packet. To accomplish this a signal is generated that indicates that a conflict has been detected, which requires the packet processor to discard all work associated with the indicated sequence number and to restart processing that packet. The processor does not have to reconstruct any information, but only to start from the beginning of the packet again assuming that all memory writes have been discarded. It must be possible to restart a packet even after packet processing has ended for that packet. This is needed because an older packet that is still processing may do something that nullifies the processing that has already been completed by a younger dependent packet.
In the present specification the term “virtual sequentiality” is used to refer to the fact that sequential processing of packets is simulated but does not actually occur. That is, results are guaranteed to be consistent with sequential processing of each packet one at a time, while in fact packets may be processed in parallel and possibly out of sequence.
The present invention is responsible for guaranteeing that the results obtained while simultaneously processing multiple packets are the same as would be obtained if all packet processing occurred in the sequence of packet arrival. This hardware mechanism is placed between the packet processing engine and the memory system as shown in
The packet processing engine 310 generates four types of commands to the HEVS mechanism 320: memory read, memory write, packet start and packet done. Each command includes a sequence number associated with the packet for which the operation applies. Returning to the packet processor are the read data returning, and a restart signal that is used to indicate to the packet processor that a particular packet needs to be restarted.
The Write Table 420 records all writes, including the sequence number, the address, the data and dependency information for forwarded reads. Each time a write is accepted from the packet processor, an entry is made in the Write Table. To save space in the Write Table, an entry can include a larger amount of storage than a single reference can generate. For example a Write Table entry could contain 64 bytes of data with valid bits associated with each byte. This would allow multiple writes to the same 64-byte area with the same sequence number to be combined into the same entry. Two different accesses with different sequence numbers cannot be combined. The Write Table is searched whenever a read is accepted to determine if data should be bypassed from the memory system. A record is kept of the fact that data has been forwarded from one packet to another so that when a packet is restarted; all packets potentially corrupted by that packet are also restarted.
When a sequence number is completed, the active packet list 430 is updated and if the oldest packet active is no longer being processed, data from the Write Table is flushed to memory and all entries for that sequence number are cleared from the Read Table. Note that the packet start command is optional in that it could be inferred from the first read or write received with a new sequence number.
In summary, the operations that are performed for each type of command from the packet processor are as follows:
Read: 1. insert entry into Read Table including sequence number and address; 2. search Write Table for the newest matching entry that has the same or an older sequence number; if one exists, forward that data back to the network processor and record dependency information; if none exists, send this request to the memory system.
Write: 1. insert entry into Write Table including sequence number, address and data; 2. search Read Table for any entry that is newer than this sequence number and matches the address; if one is found, then: a. signal restart to packet processor of the sequence number associated with the match; b. delete all Read Table entries with that sequence number; c. delete all Write Table entries with that sequence number; repeat steps a, b and c. if there are multiple matching sequence numbers, and continue down all dependency chains
Start: 1. add entry to active packet list.
Done: 1. mark packet as done in active packet list; 2. if packet is the oldest done: a. delete all Read Table entries for that sequence number; b. flush all Write Table entries for that sequence number to memory; c. delete all Write Table entries for that sequence number; d. delete the packet from the active packet list; repeat steps a. through d. if the next oldest packet is done.
However, it should be noted in this example that the read of location A by packet #2, which is the one that triggered the conflict, occurred after the write to location B by packet #2. This means that if packet #2 were able to be backed up and restarted between these two events, rather than at the beginning of processing, then packet #3 would not need to be restarted. This performance enhancement allows the amount of discarded work to be limited. It would involve a checkpoint-backup mechanism within the packet processor in conjunction with the use of sub-sequence numbers. The processor would then be able to restart processing from a safe checkpoint rather than having to back up to the beginning of the packet processing.
Another possible performance enhancement would be to save the data that is returned for each read in the Read Table. This would allow the conflict detection logic to determine if in fact wrong data had actually been returned, and only signal a restart if a data mismatch occurred. This would be advantageous if the packet processing code commonly overwrites data with the same value as is already stored, which may happen in certain applications. In addition, carefully written software could take advantage of a packet processor with this enhancement by purposefully writing an expected value into a variable early in its processing. In the common case, when the final value is written it will be the same and subsequent packets will not need to be restarted.
In many cases, it is not necessary to guarantee that processing yields the same results as if packets had been processed in the specific order of arrival, as long as the results are correct for any specific order. In other words, the packet processing code does not care which packet is processed first as long as there is a consistent order. Packets may in principle be reordered by the network in many cases and higher-level protocols cannot depend on any specific order. This principle has been illustrated in the last scenarios of
An alternative mode of operation for the packet dependency mechanism described above is to not enforce a specific order, but to signal a conflict only when no consistent order exists. In this mode the packet sequence number is being used as merely a packet identifier rather than a numeric order. The mechanism is modified so that when a write occurs, a packet sequence is defined for any previous read. Then when two packets resolve to two conflicting sequences, one must be restarted.
The present invention can be used to enforce the correctness of a sequential packet processing model in a processor that processes packets simultaneously. Ideally, there should a minimal amount of discarded work. (The importance of minimizing discarded work is particularly relevant in multithreaded processors, where multiple threads all contend for the same execution resources.) In cases where conflicts are common, more careful optimization of the restart mechanism should be implemented.
In this figure the arrow and the R represents a read in packet #2 for which a conflict arises, and the arrow and the W indicates a write in packet #1 that is the source of the conflict. In the first scenario, as soon as the conflict is detected, processing is stalled until the packet that generated the conflict has completed. This can be accomplished by having an additional stall signal to the packet processor, or by having the mechanism delay the return of the first read operation received after processing resumes for packet #2.
This first scenario is a conservative approach that essentially predicts that if a first conflict between two packets arises, then other conflicts between those two packets are likely to arise and therefore no more work should be expended on the second packet until the first packet is done.
In the second scenario of
Many more sophisticated mechanisms are possible. A conflict predictor could be implemented that would observe sequences of packet conflicts and develop a history to guess when memory operations would conflict. The information used in the predictor could be information from the packet header of each packet (e.g. source and destination IP number, and source and destination port number), as well as the memory addresses of the read and write involved. Ideally, a hardware conflict predictor would know when a read operation is received that it should not be allowed to complete since a write from an earlier packet is likely in the future. The third scenario of
Note that in the third scenario the second packet completes in the earliest time and the least machine resources are used. The advantage of a dynamic hardware conflict predictor is that it can adapt to changing application conditions and only apply stalls to read operations when there is a high likelihood of a conflict. The forth scenario of
All of the above description and teaching is specific to a single implementation of the present invention, and it should be clear to the skilled artisan that there are many alterations and amendments that might be made to the example provided, without departing from the spirit and scope of the invention.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US6389016 *||Oct 14, 1998||May 14, 2002||Nortel Networks Limited||Data communication system and method for transporting data|
|US6665755 *||Dec 22, 2000||Dec 16, 2003||Nortel Networks Limited||External memory engine selectable pipeline architecture|
|US6738379 *||Mar 30, 2000||May 18, 2004||Telefonaktiebolaget Lm Ericsson (Publ)||Method of preserving data packet sequencing|
|US20020073285 *||Sep 4, 2001||Jun 13, 2002||International Business Machines Corporation||Data sorting in information storage systems|
|US20020112100 *||May 4, 2001||Aug 15, 2002||Myron Zimmerman||System and method for data exchange|
|US20020116587 *||Dec 22, 2000||Aug 22, 2002||Modelski Richard P.||External memory engine selectable pipeline architecture|
|1||G. Sohi, S. Breach, and T. Vijaykumar, "Multiscalar processors," Proceedings of the 22nd Annual International Symposium on Computer Architecture, pp. 414-425, Ligure, Italy, Jun. 1995.|
|2||J. G. Steffan and T. Mowry, "The potential for using thread-level data speculation to facilitate automatic parallelization," Proceedings of the Fourth International Symposium on High-Performance Computer Architecture, Las Vegas, Feb. 1998.|
|3||J. Martinez and J. Torrellas, "Speculative Locks for Concurrent Execution of Critical Sections in Shared-Memory Multiprocessors," Workshop on Memory Performance Issues, International Symposium on Computer Architecture, Göteborg, Sweden , Jun. 2000.|
|4||J. Steffan, C. Colohan, Antonia Zhai, and T. Mowry, "A Scalable Approach to Thread-Level Speculation," Proceedings of the 27th Annual International Symposium on Computer Architecture, Vancouver, Canada, Jun. 2000.|
|5||L. Hammond, M. Willey, and Kunle Olukotun, "Data Speculation Support for a Chip Multiprocessor," Proceedings of the Eighth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-VIII), San Jose, Oct. 1998.|
|6||M. Cintra, J. Martinez, and J. Torrellas, "Architectural Support for Scalable Speculative Parallelization in Shared-Memory Multiprocessors," Proceedings of the 27th Annual International Symposium on Computer Architecture, Vancouver, Canada, Jun. 2000.|
|7||M. Franklin and G. Sohi, "ARB: A hardware mechanism for dynamic reordering of memory references," IEEE Transactions on Computers, vol. 45, pp. 552-571, May 1996.|
|8||M. Herlihy and J. E. B. Moss, "Transactional Memory: Architectural support for lock-free data structures," Proceedings of the International Conference on Computer Architecture, pp. 289-300, San Diego, California, May 1993.|
|9||R. Rajwar and J. Goodman, "Speculative Lock Elision: Enabling Highly Concurrent Multithreaded Execution," Proceedings of the 34th Annual International Symposiumon Microarchitecture, Austin, Texas, Dec. 2001.|
|10||S. Gopal, T. N. Vijakumar, J. E. Smith and G. S. Sohi, "Speculative versioning cache," Proceedings of the Fourth International Symposium on High-Performance Computer Architecture, Las Vegas, Feb. 1998.|
|11||S. Melvin and Y. Patt, "Handling of Packet Dependencies: A Critical Issue for Highly Parallel Network Processors," International Conference on Compilers, Architecture, and Synthesis for Embedded Systems, Oct. 8-11, 2002, Grenoble, France.|
|12||*||Tomlinson et al., "Selecting Sequence Numbers," pp 45-53, ACM, May 1974.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7441088 *||Sep 11, 2006||Oct 21, 2008||Teplin Application Limited Liability Company||Packet processor memory conflict prediction|
|US7444481||Oct 31, 2007||Oct 28, 2008||Teplin Application Limited Liability Company||Packet processor memory interface with memory conflict valve checking|
|US7475200||Oct 31, 2007||Jan 6, 2009||Teplin Application Limited Liability Company||Packet processor memory interface with write dependency list|
|US7475201||Oct 31, 2007||Jan 6, 2009||Teplin Application Limited Liability Co.||Packet processor memory interface with conditional delayed restart|
|US7478209||Oct 31, 2007||Jan 13, 2009||Teplin Application Limited Liability Co.||Packet processor memory interface with conflict detection and checkpoint repair|
|US7487304||Oct 31, 2007||Feb 3, 2009||Teplin Application Limited||Packet processor memory interface with active packet list|
|US7496721||Oct 31, 2007||Feb 24, 2009||Teplin Application Limited||Packet processor memory interface with late order binding|
|US7506104 *||Oct 31, 2007||Mar 17, 2009||Teplin Application Limited Liability Company||Packet processor memory interface with speculative memory reads|
|US7739452||Mar 9, 2009||Jun 15, 2010||Stephen Waller Melvin||Method and apparatus for hardware enforced virtual sequentiality|
|US7765405 *||Jul 27, 2010||Microsoft Corporation||Receive side scaling with cryptographically secure hashing|
|US7808999 *||Mar 24, 2008||Oct 5, 2010||Ericsson Ab||Method and apparatus for out-of-order processing of packets using linked lists|
|US7852846||Dec 14, 2010||Ericsson Ab||Method and apparatus for out-of-order processing of packets|
|US8095741||Jan 10, 2012||International Business Machines Corporation||Transactional memory computing system with support for chained transactions|
|US8095750||Oct 30, 2007||Jan 10, 2012||International Business Machines Corporation||Transactional memory system with fast processing of common conflicts|
|US8117403||Oct 30, 2007||Feb 14, 2012||International Business Machines Corporation||Transactional memory system which employs thread assists using address history tables|
|US8145903 *||May 25, 2007||Mar 27, 2012||Red Hat, Inc.||Method and system for a kernel lock validator|
|US8321637||Nov 27, 2012||International Business Machines Corporation||Computing system with optimized support for transactional memory|
|US8566524||Aug 31, 2009||Oct 22, 2013||International Business Machines Corporation||Transactional memory system with efficient cache support|
|US8667231||Oct 22, 2012||Mar 4, 2014||International Business Machines Corporation||Transactional memory system with efficient cache support|
|US8688920||Oct 30, 2007||Apr 1, 2014||International Business Machines Corporation||Computing system with guest code support of transactional memory|
|US8738862||Oct 23, 2013||May 27, 2014||International Business Machines Corporation||Transactional memory system with efficient cache support|
|US8774177 *||Jul 17, 2002||Jul 8, 2014||Alcatel Lucent||Classifying traffic at a network node using multiple on-chip memory arrays|
|US9009452||Oct 30, 2007||Apr 14, 2015||International Business Machines Corporation||Computing system with transactional memory using millicode assists|
|US9104427||Oct 19, 2012||Aug 11, 2015||International Business Machines Corporation||Computing system with transactional memory using millicode assists|
|US20030223423 *||Jul 17, 2002||Dec 4, 2003||James Yu||Classifying traffic at a network node using multiple on-chip memory arrays|
|US20050243734 *||Jun 14, 2005||Nov 3, 2005||Nemirovsky Mario D||Multi-threaded packet processing engine for stateful packet processing|
|US20060195698 *||Feb 25, 2005||Aug 31, 2006||Microsoft Corporation||Receive side scaling with cryptographically secure hashing|
|US20060218556 *||May 17, 2005||Sep 28, 2006||Nemirovsky Mario D||Mechanism for managing resource locking in a multi-threaded environment|
|US20080259928 *||Mar 24, 2008||Oct 23, 2008||Chen Edmund G||Method and Apparatus for Out-of-Order Processing of Packets using Linked Lists|
|US20080259960 *||Mar 24, 2008||Oct 23, 2008||Favor John G||Method and apparatus for Out-of-Order Processing of Packets|
|US20080288726 *||Oct 30, 2007||Nov 20, 2008||International Business Machines Corporation||Transactional Memory System with Fast Processing of Common Conflicts|
|US20080288727 *||May 14, 2007||Nov 20, 2008||International Business Machines Corporation||Computing System with Optimized Support for Transactional Memory|
|US20080288730 *||Oct 30, 2007||Nov 20, 2008||International Business Machines Corporation||Transactional Memory System Which Employs Thread Assists Using Address History Tables|
|US20080294892 *||May 25, 2007||Nov 27, 2008||Ingo Molnar||Method and system for a kernel lock validator|
|US20090113443 *||Oct 30, 2007||Apr 30, 2009||International Business Machines Corporation||Transactional Memory Computing System with Support for Chained Transactions|
|US20090193216 *||Mar 9, 2009||Jul 30, 2009||Stephen Melvin||Method and Apparatus for Hardware Enforced Virtual Sequentiality|
|US20100202292 *||Aug 12, 2010||Nemirovsky Mario D||Mechanism for Achieving Packet Flow Control In a Multi-Threaded, Multi-Packet Environment|
|US20100205608 *||Feb 2, 2010||Aug 12, 2010||Nemirovsky Mario D||Mechanism for Managing Resource Locking in a Multi-Threaded Environment|
|US20110055483 *||Aug 31, 2009||Mar 3, 2011||International Business Machines Corporation||Transactional memory system with efficient cache support|
|U.S. Classification||711/154, 370/394, 712/218, 709/229, 712/225, 711/221, 711/158, 709/216|
|International Classification||G06F12/00, H04L12/56|
|Cooperative Classification||H04L49/9094, H04L49/90|
|European Classification||H04L49/90S, H04L49/90|
|Mar 13, 2007||CC||Certificate of correction|
|Mar 2, 2008||AS||Assignment|
Owner name: TEPLIN APPLICATION LIMITED, DELAWARE
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MELVIN, STEPHEN;REEL/FRAME:020582/0432
Effective date: 20071222
|Apr 2, 2008||AS||Assignment|
Owner name: TEPLIN APPLICATION LIMITED LIABILITY COMPANY, DELA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MELVIN, STEPHEN;REEL/FRAME:020783/0780
Effective date: 20071222
|May 21, 2009||FPAY||Fee payment|
Year of fee payment: 4
|Mar 18, 2013||FPAY||Fee payment|
Year of fee payment: 8
|Dec 30, 2015||AS||Assignment|
Owner name: F. POSZAT HU, L.L.C., DELAWARE
Free format text: MERGER;ASSIGNOR:TEPLIN APPLICATION LIMITED LIABILITY COMPANY;REEL/FRAME:037387/0264
Effective date: 20150812